# Arena synthesis note: Prime Academy strategy parameters and the rule that picks them

Date 2026-09-28. Three candidates, one per model family: candidate-1 Opus, candidate-2 Fable,
candidate-3 GPT through Codex. Cross-judge: GPT through Codex, read-only, on the rubric in
`judge/RUBRIC.md`. No dropouts (candidate-3 needed a relaunch with `--skip-git-repo-check`;
its first attempt wrote nothing).

## Convergence

All three chose the same core: reward-to-risk 1 in both stages, 3% risk in the assessment, 1.5%
once funded, the $3,000 tier, EURUSD short, entries 07:00-10:00 UTC, the deployed stop, the target
clipped, the stop at the floor, stepping aside for news. That is a strong agreement signal, and
it moves the assessment from the notebook's 4% to 3% for one reason all three gave: a 4% win is
44% of the 9% target, and whether the 35% rule applies in the assessment is unverified.

They diverged on how many accounts and when to buy: candidate-1 two accounts on day one ($76),
candidate-2 one account behind a quantitative purchase gate, candidate-3 zero purchases until
economic and execution gates are met.

## Pick

**Base: candidate-2.** The judge recommended candidate-3. The disagreement is on criterion 4,
actionable: candidate-3's rule is a paragraph of qualitative gates, candidate-2's is one
quantitative gate (expected cash per fee at least 1.33 at the measured cost, which at a 0.5-pip
spread means at most 0.55 step-asides a trade) plus the same settings. Under the measured 1.33
asides a trade candidate-2's gate is closed, so it reaches candidate-3's conclusion, zero
purchases now, by a rule the operator can re-apply with two numbers. On criterion 5 candidate-2's
declared test is on the resolvable stage with stated power; candidate-3's is on the funded stage,
which it concedes the history cannot resolve. On criterion 2 the judge is right that candidate-2's
daily-loss check is from entry, not from the previous day's equity; that is grafted from
candidate-1 below. The judge's ranking of candidate-3 on evidence discipline is accepted and its
framing is grafted.

The maintainer's test: candidate-2's rule has the smaller interface and the cleaner boundary
between the settings, the gate and the test.

## Grafts

From candidate-1:
- **The strict daily-loss reading.** rules.json makes the daily limit 5% of the previous day's
  equity, equity-based. A trade held across the day boundary while in profit raises that
  reference, so a stop-out the next day can lose more than 5% from it. At 3% risk the breach needs
  an overnight floating gain above 0.64 of the stop distance; at 4%, 0.22; at 1.5%, 2.3. The
  broker scores accounts only at completed trades and cannot see this. Grafted as a rule reading
  to test, and as a reason 3% is safer than 4% beyond the 35% cap.
- **The spread sensitivity of the funded risk:** 1.5% holds from 0.25 to 1.0 pip; at zero spread
  1% takes over. Grafted into "what would change the choice".
- Rejected: buying two accounts on day one, and the power claim of its strict-daily test, which
  transfers sample sizes from a different statistic.

From candidate-3:
- **The evidence-discipline framing:** the news cost is a sensitivity computed by transferring a
  rate measured at another cell, not a measurement at this one; the broker's unresolvable cells
  are neither validated nor refuted; the investor's identity is a prior, not data; a purchase is
  released by gates, and a failed gate means zero, not a larger bet.
- **Execution controls before any purchase:** floating-equity daily check, calendar-day payout
  clock, no Friday-close order that fills Sunday.
- Rejected: its funded-stage falsification test, which cannot reach its declared power on this
  history; candidate-2's assessment-stage test is kept instead.

From candidate-2 itself, kept as written: the settings, the gate, the honour-rate sensitivity
(wallet log growth +0.167 at P(paid) 1.0, near zero at 0.8, negative at 0.6), the fallback of
2.5% if the target clip is not wired live, and the declared test on the assessment with its
power.

## Verification

Every number the synthesis cites was recomputed by the picker from `modules/prime.py`:
assessment 3% clipped 45.38% (4%: 46.24%; unclipped 3%: 40.19%); funded 1.5% any payout
52.36%, 1.10 payouts, 6.66% withdrawn, first payout median 18 trades; per purchase P(cash)
23.8%, expected cash $63.45 against the $38 fee at a 2.87% cost share; $36.81 at 6.7%; $59.40 at
3.4% (asides charged commission only). The 35% arithmetic: a 3% win is 2.91% of the bankroll
against a 3.15% ceiling; a 4% win is 3.89%. The strict-daily thresholds: 0.64, 0.22 and 2.3 of
the stop at 3%, 4% and 1.5%. All agree with the candidates.

Not verified, because they are the next tests: the aside rate on a EUR+USD-only calendar, the
real spread at the aside instants, and the strict-daily breach count. Those are the three
measurements that decide whether the gate opens.